Medical vision-language models (VLMs) for visual question answering typically pass long sequences of visual tokens to the language model. This makes token pruning an attractive way to reduce computational load, but it is also risky for medical images, where the correct answer may depend on information that pruning could easily discard.
The new method, MedSTP, is proposed to address this tension by preserving what the abstract calls "subspace and transition evidence" during pruning. Rather than simply dropping tokens based on salience, it aims to keep the structural and relational cues that are important for answering medical questions.
This work highlights a key challenge in efficient medical VQA: aggressive pruning can remove the very evidence needed for accurate answers. MedSTP is presented as a way to navigate that trade-off, though the abstract does not yet include quantitative results or comparisons with other pruning approaches.